diff --git a/docs/user_guide/examples/offline_inference/mammothmodal2_preview.md b/docs/user_guide/examples/offline_inference/mammothmodal2_preview.md index 6ab31298c..60f1cbe5f 100644 --- a/docs/user_guide/examples/offline_inference/mammothmodal2_preview.md +++ b/docs/user_guide/examples/offline_inference/mammothmodal2_preview.md @@ -22,7 +22,7 @@ Text-to-image now runs through the shared offline image example ```bash python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \ --model ./MammothModa2-Preview \ - --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \ + --deploy-config ./vllm_omni/deploy/mammoth_moda2_ar.yaml \ --question "Summarize this image." \ --image ./image.png ``` diff --git a/examples/offline_inference/mammothmodal2_preview/README.md b/examples/offline_inference/mammothmodal2_preview/README.md index 16937680a..bb68f3845 100644 --- a/examples/offline_inference/mammothmodal2_preview/README.md +++ b/examples/offline_inference/mammothmodal2_preview/README.md @@ -19,7 +19,7 @@ hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-P ```bash python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \ --model ./MammothModa2-Preview \ - --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \ + --deploy-config ./vllm_omni/deploy/mammoth_moda2_ar.yaml \ --question "Summarize this image." \ --image ./image.png ``` diff --git a/examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py b/examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py index ca87b9e9a..7a22803c9 100644 --- a/examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py +++ b/examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py @@ -4,7 +4,7 @@ Offline inference example: MammothModa2 image summarization (single AR stage). Example: uv run python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \ --model path/to/MammothModa2-Preview \ - --stage-config vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \ + --deploy-config vllm_omni/deploy/mammoth_moda2_ar.yaml \ --image /path/to/input.jpg \ --question "Please summarize the content of this image." """ @@ -27,9 +27,7 @@ DEFAULT_QUESTION = "Please summarize the content of this image." def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser(description="MammothModa2 image summarization (offline, AR only).") parser.add_argument("--model", type=str, required=True, help="Path to model directory or model id.") - parser.add_argument( - "--stage-config", type=str, required=True, help="Path to stage config yaml (single-stage AR->text)." - ) + parser.add_argument("--deploy-config", type=str, required=True, help="Path to deploy config yaml.") parser.add_argument("--image", type=str, required=True, help="Path to input image.") parser.add_argument("--question", type=str, default=DEFAULT_QUESTION, help="Question/instruction for the model.") parser.add_argument("--system", type=str, default=DEFAULT_SYSTEM, help="System prompt.") @@ -73,7 +71,7 @@ def main() -> None: omni = Omni( model=args.model, - stage_configs_path=args.stage_config, + deploy_config=args.deploy_config, trust_remote_code=args.trust_remote_code, enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler, ) diff --git a/examples/offline_inference/text_to_speech/omnivoice/end2end.py b/examples/offline_inference/text_to_speech/omnivoice/end2end.py index 619f8160b..a6bdc7b4e 100644 --- a/examples/offline_inference/text_to_speech/omnivoice/end2end.py +++ b/examples/offline_inference/text_to_speech/omnivoice/end2end.py @@ -34,7 +34,7 @@ def run_e2e(): help="Model name or path (HuggingFace or local)", ) parser.add_argument( - "--stage-config", + "--deploy-config", type=str, default="vllm_omni/deploy/omnivoice.yaml", ) @@ -87,14 +87,14 @@ def run_e2e(): ) args = parser.parse_args() - if not os.path.exists(args.stage_config): - raise FileNotFoundError(f"Stage config not found: {args.stage_config}") + if not os.path.exists(args.deploy_config): + raise FileNotFoundError(f"Deploy config not found: {args.deploy_config}") print(f"Initializing OmniVoice with model={args.model}") omni = Omni( model=args.model, - stage_configs_path=args.stage_config, + deploy_config=args.deploy_config, log_stats=True, ) diff --git a/recipes/MammothModa2/MammothModa2-Preview.md b/recipes/MammothModa2/MammothModa2-Preview.md index 289efece0..0733b0630 100644 --- a/recipes/MammothModa2/MammothModa2-Preview.md +++ b/recipes/MammothModa2/MammothModa2-Preview.md @@ -32,12 +32,12 @@ flags. Image size uses the standard `--height` / `--width` flags. [`examples/offline_inference/text_to_image/text_to_image.py`](../../examples/offline_inference/text_to_image/text_to_image.py) - Declared parameters: [`vllm_omni/model_extras/mammothmodal2_preview.py`](../../vllm_omni/model_extras/mammothmodal2_preview.py) -- Stage config: - [`vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml`](../../vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml) +- Deploy config: + [`vllm_omni/deploy/mammoth_moda2.yaml`](../../vllm_omni/deploy/mammoth_moda2.yaml) ## Hardware Support -The default stage config runs both the AR and DiT stages on a single GPU +The default deploy config runs both the AR and DiT stages on a single GPU (`devices: "0"`). The committed `gpu_memory_utilization` split (stage-0 AR `0.5`, stage-1 DiT `0.3`) is sized for an ~80 GB GPU. The model also fits on a 48 GB GPU after rebalancing the split so the AR weights (~23 GB) leave room for the KV @@ -48,7 +48,7 @@ cache — see the note under *1x L40S 48GB*. ### 1x L40S 48GB > **48 GB config adjustment:** the committed -> `vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml` uses +> `vllm_omni/deploy/mammoth_moda2.yaml` uses > `gpu_memory_utilization` 0.5 / 0.3 (sized for ~80 GB). To fit on a 48 GB L40S, > set the stage-0 (AR) value to `0.8` and the stage-1 (DiT) value to `0.16` > before running. (On an ~80 GB GPU, leave the defaults unchanged.) @@ -71,13 +71,13 @@ hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-P ``` Run text-to-image with the shared offline example from the repository root. The -stage config sets `trust_remote_code`, so no extra flag is needed. Forward the +deploy config sets `trust_remote_code`, so no extra flag is needed. Forward the MammothModa2 generation parameters as a JSON object through `--extra-body`: ```bash python examples/offline_inference/text_to_image/text_to_image.py \ --model ./MammothModa2-Preview \ - --stage-configs-path vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml \ + --stage-configs-path vllm_omni/deploy/mammoth_moda2.yaml \ --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \ --height 1024 \ --width 1024 \ diff --git a/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py b/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py index b0d82dbc4..04308c4a7 100644 --- a/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py +++ b/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py @@ -5,7 +5,7 @@ Verifies that the AR->DiT pipeline produces an image tensor whose pixel values match a golden reference. Model Hub repo id: ``bytedance-research/MammothModa2-Preview``. -Stage config: ``get_deploy_config_path("mammoth_moda2.yaml")`` → ``vllm_omni/deploy/mammoth_moda2.yaml`` +Deploy config: ``get_deploy_config_path("mammoth_moda2.yaml")`` -> ``vllm_omni/deploy/mammoth_moda2.yaml`` Golden pixel file: ``tests/e2e/offline_inference/fixtures/mammoth_moda2_t2i_golden.json`` Regenerate with: ``UPDATE_GOLDEN=1 pytest tests/e2e/offline_inference/test_mammoth_moda2_expansion.py`` @@ -36,9 +36,9 @@ _VISION_END_TOKEN_ID = 151653 # "<|vision_end|>" _AR_PATCH_SIZE = 16 MODEL_PATH = "bytedance-research/MammothModa2-Preview" -T2I_STAGE_CONFIG = get_deploy_config_path("mammoth_moda2.yaml") +T2I_DEPLOY_CONFIG = get_deploy_config_path("mammoth_moda2.yaml") -_OMNI_RUNNER_PARAM = (MODEL_PATH, T2I_STAGE_CONFIG) +_OMNI_RUNNER_PARAM = (MODEL_PATH, None, {"deploy_config": T2I_DEPLOY_CONFIG}) # Golden pixel reference file. Set UPDATE_GOLDEN=1 to regenerate. _GOLDEN_T2I_PATH = Path(__file__).parent / "fixtures" / "mammoth_moda2_t2i_golden.json" diff --git a/tests/e2e/offline_inference/test_omnivoice_expansion.py b/tests/e2e/offline_inference/test_omnivoice_expansion.py index 692278ef7..e3a138112 100644 --- a/tests/e2e/offline_inference/test_omnivoice_expansion.py +++ b/tests/e2e/offline_inference/test_omnivoice_expansion.py @@ -18,13 +18,15 @@ from tests.helpers.runtime import OmniRunner from tests.helpers.stage_config import get_deploy_config_path MODEL = "k2-fsa/OmniVoice" -STAGE_CONFIG = get_deploy_config_path("omnivoice.yaml") +DEPLOY_CONFIG = get_deploy_config_path("omnivoice.yaml") -# (model, stage_config_path, extra_omni_kwargs) — see ``omni_runner`` in tests.helpers.fixtures.runtime +# OmniRunner tuple: model, legacy stage config path, extra Omni kwargs. +# The migrated test passes deploy_config through extra Omni kwargs. _OMNI_RUNNER_PARAM = ( MODEL, - STAGE_CONFIG, + None, { + "deploy_config": DEPLOY_CONFIG, "trust_remote_code": True, "log_stats": True, }, diff --git a/tests/test_config_factory.py b/tests/test_config_factory.py index 596a13dbb..717b9a5d7 100644 --- a/tests/test_config_factory.py +++ b/tests/test_config_factory.py @@ -749,6 +749,32 @@ stages: runtime_config = deploy.stages[0].engine_extras["hf_overrides"]["voxcpm2_runtime_config"] assert runtime_config == expected_runtime_config + @pytest.mark.parametrize( + ("deploy_name", "pipeline_name", "stage_count", "final_output_type"), + [ + ("mammoth_moda2.yaml", "mammoth_moda2", 2, "image"), + ("mammoth_moda2_ar.yaml", "mammoth_moda2_ar", 1, "text"), + ("omnivoice.yaml", "omnivoice", 1, "audio"), + ], + ) + def test_load_new_registry_backed_deploy_configs( + self, + deploy_name: str, + pipeline_name: str, + stage_count: int, + final_output_type: str, + ): + deploy_path = Path(get_deploy_config_path(deploy_name)) + deploy = load_deploy_config(deploy_path) + assert deploy.pipeline == pipeline_name + + with patch("vllm_omni.platforms.current_omni_platform") as platform: + platform.device_name = "cuda" + stages = merge_pipeline_deploy(OMNI_PIPELINES[pipeline_name], deploy) + assert len(stages) == stage_count + assert stages[-1].final_output is True + assert stages[-1].final_output_type == final_output_type + def test_merge_pipeline_deploy(self): deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml" if not deploy_path.exists(): diff --git a/vllm_omni/config/pipeline_registry.py b/vllm_omni/config/pipeline_registry.py index 7bf79d44c..934d15d18 100644 --- a/vllm_omni/config/pipeline_registry.py +++ b/vllm_omni/config/pipeline_registry.py @@ -56,6 +56,10 @@ from vllm_omni.model_executor.models.hunyuan_image3.pipeline import ( ) from vllm_omni.model_executor.models.indextts2.pipeline import INDEXTTS2_PIPELINE from vllm_omni.model_executor.models.lance.pipeline import LANCE_PIPELINE +from vllm_omni.model_executor.models.mammoth_moda2.pipeline import ( + MAMMOTH_MODA2_AR_PIPELINE, + MAMMOTH_MODA2_PIPELINE, +) from vllm_omni.model_executor.models.mimo_audio.pipeline import MIMO_AUDIO_PIPELINE from vllm_omni.model_executor.models.ming_flash_omni.pipeline import ( MING_FLASH_OMNI_IMAGE_PIPELINE, @@ -74,6 +78,7 @@ from vllm_omni.model_executor.models.moss_tts.pipeline import ( MOSS_TTS_REALTIME_PIPELINE, ) from vllm_omni.model_executor.models.moss_tts_nano.pipeline import MOSS_TTS_NANO_PIPELINE +from vllm_omni.model_executor.models.omnivoice.pipeline import OMNIVOICE_PIPELINE from vllm_omni.model_executor.models.qwen2_5_omni.pipeline import ( QWEN2_5_OMNI_PIPELINE, QWEN2_5_OMNI_THINKER_ONLY_PIPELINE, @@ -118,6 +123,9 @@ OMNI_PIPELINES: dict[str, PipelineConfig | PipelineResolverFunc] = { "ming_flash_omni_thinker_only": MING_FLASH_OMNI_THINKER_ONLY_PIPELINE, "ming_flash_omni_image": MING_FLASH_OMNI_IMAGE_PIPELINE, "moss_tts_nano": MOSS_TTS_NANO_PIPELINE, + "omnivoice": OMNIVOICE_PIPELINE, + "mammoth_moda2": MAMMOTH_MODA2_PIPELINE, + "mammoth_moda2_ar": MAMMOTH_MODA2_AR_PIPELINE, "moss_tts_delay": MOSS_TTS_PIPELINE, "moss_tts_realtime": MOSS_TTS_REALTIME_PIPELINE, "moss_tts_local": MOSS_TTS_LOCAL_PIPELINE, diff --git a/vllm_omni/deploy/mammoth_moda2.yaml b/vllm_omni/deploy/mammoth_moda2.yaml index d1a49d662..afae38ee0 100644 --- a/vllm_omni/deploy/mammoth_moda2.yaml +++ b/vllm_omni/deploy/mammoth_moda2.yaml @@ -1,42 +1,27 @@ -# MammothModa2 AR→DiT (text-to-image). Canonical copy lives at -# ``vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml``; keep in sync. -# -# Must remain legacy ``stage_args`` layout (no top-level ``stages``): Omni resolves -# explicit ``--stage-configs-path`` via ``load_stage_configs_from_yaml``. +# MammothModa2 AR→DiT (text-to-image). -stage_args: +async_chunk: false +pipeline: mammoth_moda2 + +stages: - stage_id: 0 - runtime: - devices: "0" - engine_args: - model_stage: ar - max_num_seqs: 100 - model_arch: MammothModa2ForConditionalGeneration - worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker - scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler - max_model_len: 8192 - gpu_memory_utilization: 0.5 - enforce_eager: true - trust_remote_code: true - engine_output_type: latent - enable_prefix_caching: false - final_output: false + devices: "0" + max_num_seqs: 100 + max_model_len: 8192 + gpu_memory_utilization: 0.5 + enforce_eager: true + trust_remote_code: true + enable_prefix_caching: false - stage_id: 1 - runtime: - devices: "0" - engine_args: - model_stage: dit - max_num_seqs: 1 - model_arch: MammothModa2ForConditionalGeneration - worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker - scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler - gpu_memory_utilization: 0.3 - enforce_eager: true - trust_remote_code: true - engine_output_type: image - enable_prefix_caching: false - engine_input_source: [0] - custom_process_input_func: vllm_omni.model_executor.stage_input_processors.mammoth_moda2.ar2dit - final_output: true - final_output_type: image + devices: "0" + max_num_seqs: 1 + gpu_memory_utilization: 0.3 + enforce_eager: true + trust_remote_code: true + enable_prefix_caching: false + default_sampling_params: + extra_args: + text_guidance_scale: 9.0 + cfg_range: [0.0, 1.0] + num_inference_steps: 50 diff --git a/vllm_omni/deploy/mammoth_moda2_ar.yaml b/vllm_omni/deploy/mammoth_moda2_ar.yaml new file mode 100644 index 000000000..45e2580d9 --- /dev/null +++ b/vllm_omni/deploy/mammoth_moda2_ar.yaml @@ -0,0 +1,14 @@ +# MammothModa2 AR-only (image/video understanding and summarization). + +async_chunk: false +pipeline: mammoth_moda2_ar + +stages: + - stage_id: 0 + devices: "0" + max_num_seqs: 16 + max_model_len: 8192 + gpu_memory_utilization: 0.5 + enforce_eager: true + trust_remote_code: true + enable_prefix_caching: false diff --git a/vllm_omni/deploy/omnivoice.yaml b/vllm_omni/deploy/omnivoice.yaml index 6491be6dc..6769dd5e9 100644 --- a/vllm_omni/deploy/omnivoice.yaml +++ b/vllm_omni/deploy/omnivoice.yaml @@ -1,19 +1,13 @@ # OmniVoice TTS deploy: diffusion engine single stage (dit). # Text → iterative unmasking → codebook tokens → DAC decode → 24kHz audio. -# (Migrated from model_executor/stage_configs/omnivoice.yaml for deploy/ layout.) -stage_args: +async_chunk: false +pipeline: omnivoice + +stages: - stage_id: 0 - stage_type: diffusion - is_comprehension: true - runtime: - devices: 0 - engine_args: - model_stage: dit - model_class_name: "OmniVoicePipeline" - enforce_eager: true - trust_remote_code: true - distributed_executor_backend: "mp" - dtype: "float32" - final_output: true - final_output_type: audio + devices: "0" + enforce_eager: true + trust_remote_code: true + distributed_executor_backend: "mp" + dtype: "float32" diff --git a/vllm_omni/diffusion/models/mammoth_moda2/pipeline_mammothmoda2_dit.py b/vllm_omni/diffusion/models/mammoth_moda2/pipeline_mammothmoda2_dit.py index 0c312aa7e..614c8268c 100644 --- a/vllm_omni/diffusion/models/mammoth_moda2/pipeline_mammothmoda2_dit.py +++ b/vllm_omni/diffusion/models/mammoth_moda2/pipeline_mammothmoda2_dit.py @@ -9,6 +9,7 @@ from diffusers.utils.torch_utils import randn_tensor from torch import nn from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm from vllm.config import VllmConfig +from vllm.logger import init_logger from vllm.model_executor.models.utils import AutoWeightsLoader, WeightsMapper from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery @@ -19,6 +20,8 @@ from .mammothmoda2_dit_model import SimpleQFormerImageRefiner, Transformer2DMode from .rope_real import RotaryPosEmbedReal from .schedulers import FlowMatchEulerDiscreteScheduler +logger = init_logger(__name__) + class MammothModa2DiTPipeline(nn.Module, SupportsComponentDiscovery): """ @@ -60,14 +63,14 @@ class MammothModa2DiTPipeline(nn.Module, SupportsComponentDiscovery): self.gen_transformer = Transformer2DModel.from_config(self.config.gen_dit_config) # llm_config is a Mammothmoda2Qwen2_5_VLConfig which has nested text_config - llm_hidden_size = int( - getattr( - getattr(self.config.llm_config, "text_config", None), - "hidden_size", - 0, - ) - or 0 - ) + llm_hidden_size = 0 + text_config = self.config.get_text_config() + if text_config is None: + logger.warning("No text config; failed to infer llm_hidden_size.") + elif not hasattr(text_config, "hidden_size"): + logger.warning("Text config exists, but has no hidden_size attribute; failed to infer llm_hidden_size.") + else: + llm_hidden_size = int(text_config.hidden_size or 0) if llm_hidden_size <= 0: raise ValueError( "Failed to infer llm hidden_size from Mammothmoda2Config.llm_config.text_config.hidden_size" diff --git a/vllm_omni/model_executor/models/mammoth_moda2/pipeline.py b/vllm_omni/model_executor/models/mammoth_moda2/pipeline.py new file mode 100644 index 000000000..be35eb562 --- /dev/null +++ b/vllm_omni/model_executor/models/mammoth_moda2/pipeline.py @@ -0,0 +1,66 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""MammothModa2 pipeline topology (frozen). + +Stage 0: AR — multimodal understanding + latent generation +Stage 1: DiT — latent → image + +For text/image understanding tasks (text output only), use MAMMOTH_MODA2_AR_PIPELINE. +""" + +from vllm_omni.config.stage_config import ( + PipelineConfig, + StageExecutionType, + StagePipelineConfig, +) + +_PROC = "vllm_omni.model_executor.stage_input_processors.mammoth_moda2" + +MAMMOTH_MODA2_PIPELINE = PipelineConfig( + model_type="mammoth_moda2", + model_arch="MammothModa2ForConditionalGeneration", + hf_architectures=("Mammothmoda2Model", "MammothModa2ForConditionalGeneration"), + stages=( + StagePipelineConfig( + stage_id=0, + model_stage="ar", + execution_type=StageExecutionType.LLM_AR, + input_sources=(), + final_output=False, + owns_tokenizer=True, + requires_multimodal_data=True, + engine_output_type="latent", + ), + StagePipelineConfig( + stage_id=1, + model_stage="dit", + execution_type=StageExecutionType.LLM_GENERATION, + input_sources=(0,), + final_output=True, + final_output_type="image", + owns_tokenizer=False, + requires_multimodal_data=False, + engine_output_type="image", + custom_process_input_func=f"{_PROC}.ar2dit", + ), + ), +) + +# Single-stage AR variant for understanding / summarization tasks +MAMMOTH_MODA2_AR_PIPELINE = PipelineConfig( + model_type="mammoth_moda2_ar", + model_arch="MammothModa2ForConditionalGeneration", + stages=( + StagePipelineConfig( + stage_id=0, + model_stage="ar", + execution_type=StageExecutionType.LLM_AR, + input_sources=(), + final_output=True, + final_output_type="text", + owns_tokenizer=True, + requires_multimodal_data=True, + engine_output_type="text", + ), + ), +) diff --git a/vllm_omni/model_executor/models/omnivoice/pipeline.py b/vllm_omni/model_executor/models/omnivoice/pipeline.py new file mode 100644 index 000000000..88d479459 --- /dev/null +++ b/vllm_omni/model_executor/models/omnivoice/pipeline.py @@ -0,0 +1,31 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""OmniVoice pipeline topology (frozen). + +Single-stage diffusion model for text-to-speech. +Stage 0: DiT — text → iterative unmasking → codebook tokens → DAC decode → 24kHz audio +""" + +from vllm_omni.config.stage_config import ( + PipelineConfig, + StageExecutionType, + StagePipelineConfig, +) + +OMNIVOICE_PIPELINE = PipelineConfig( + model_type="omnivoice", + model_arch="OmniVoicePipeline", + stages=( + StagePipelineConfig( + stage_id=0, + model_stage="dit", + execution_type=StageExecutionType.DIFFUSION, + input_sources=(), + final_output=True, + final_output_type="audio", + owns_tokenizer=False, + requires_multimodal_data=False, + engine_output_type="audio", + ), + ), +) diff --git a/vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml b/vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml deleted file mode 100644 index f5e449253..000000000 --- a/vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml +++ /dev/null @@ -1,51 +0,0 @@ -# Stage config for running MammothModa2 with full AR→DiT pipeline (text-to-image). -# Stage 0: AR (multimodal understanding + latent generation) -# Stage 1: DiT (latent → image) -# -# For text/image understanding tasks (text output only), use mammoth_moda2_ar.yaml. - -stage_args: - - stage_id: 0 - runtime: - devices: "0" - engine_args: - model_stage: ar - max_num_seqs: 100 - model_arch: MammothModa2ForConditionalGeneration - worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker - scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler - max_model_len: 8192 - gpu_memory_utilization: 0.5 - enforce_eager: true - trust_remote_code: true - engine_output_type: latent - enable_prefix_caching: false - final_output: false - - - stage_id: 1 - runtime: - devices: "0" - engine_args: - model_stage: dit - max_num_seqs: 1 - model_arch: MammothModa2ForConditionalGeneration - worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker - scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler - gpu_memory_utilization: 0.3 - enforce_eager: true - trust_remote_code: true - engine_output_type: image - enable_prefix_caching: false - # Declares extra_args so has_sampling_extra_args is enabled, routing the t2i knobs - # (text_guidance_scale, cfg_range, num_inference_steps) declared in - # vllm_omni/model_extras/mammothmodal2_preview.py from extra_body into the DiT - # forward. Values here are the defaults applied when a request omits a knob. - default_sampling_params: - extra_args: - text_guidance_scale: 9.0 - cfg_range: [0.0, 1.0] - num_inference_steps: 50 - engine_input_source: [0] - custom_process_input_func: vllm_omni.model_executor.stage_input_processors.mammoth_moda2.ar2dit - final_output: true - final_output_type: image diff --git a/vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml b/vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml deleted file mode 100644 index d455c7973..000000000 --- a/vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml +++ /dev/null @@ -1,24 +0,0 @@ -# Stage config for running MammothModa2 with single AR stage (understanding / summarization). -# Stage 0: AR (multimodal understanding + text generation) -# -# Use this config for image/video understanding tasks where only text output is needed. -# For text-to-image generation (AR → DiT), use mammoth_moda2.yaml instead. - -stage_args: - - stage_id: 0 - runtime: - devices: "0" - engine_args: - model_stage: ar - max_num_seqs: 16 - model_arch: MammothModa2ForConditionalGeneration - worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker - scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler - max_model_len: 8192 - gpu_memory_utilization: 0.5 - enforce_eager: true - trust_remote_code: true - engine_output_type: text - enable_prefix_caching: false - final_output: true - final_output_type: text